Lore

Unsorted

Developments

2026-10-15

In "Pro Photographer Ranks AI Image Generators (Late 2026)", the photographer Bernlieger frames generative image AI as a difficulty-and-value pyramid. Generic prompts at the bottom, like a dune in the desert, have a million right answers, so every model passes and the output carries little value. Each added detail narrows the acceptable outputs, which raises both difficulty and perceived value. They test GPT 2.5, Nano Banana 2, Grok Imagine 2 and Midjourney V8 on eight scenarios, using the same prompt and often a deliberately poor iPhone source for every model, and they warn that at September 2026 speed the ranking could look different within months. At the low end the rankings follow taste and scatter. Grok wins the Alps convertible and is last on the beach photo, and GPT does the reverse. The faults flagged are plausibility errors such as driving side, footprints, finger anatomy and a third child. Once a real object or photo must be reproduced, fidelity outranks beauty. Grok makes the best-looking watch ad but alters the dial, GPT reproduces the watch exactly, and Midjourney's stunning scenery fails the main idea of keeping the object's visual integrity. Nano Banana is the exception on the selfie-to-portrait task, because it is willing to change hair, lighting and contrast. That same willingness hurts it on preservation edits, where it alters pose, crop and sky. At the top (people removal, summer-to-winter, interior retouching) the winning behaviour is judgment and restraint: change only what was asked. GPT leads all three, for example by removing the museum's security check while keeping everything else, and the speaker admits this makes them nervous about their job as a hospitality photographer. Grok over-removes and invents content, and Midjourney needs its manual paint-over tool. Errors are weighed by how fixable they are in post: a scratched dial can be repaired, but a changed building cannot. Scored 4/3/2/1 per scenario, GPT-2.5 is the clear winner and never placed below third. Grok and Nano Banana sit close together in the middle, and Midjourney is last. The recommendation follows the pyramid. Any model works at the bottom, where Midjourney's cinematic, less-AI look is an advantage. Higher up, accuracy matters, so Grok and especially GPT-2.5 are the picks. A final 9:15 analog clock test supports this: Midjourney looks cinematic but shows the wrong time, Grok draws three hands, and only GPT-2.5 consistently shows the right time, so correct time on a clock or watch is still a weak spot for AI. The speaker also notes that careful inputs, such as a better-prepared selfie, would lift every result, so the deliberately poor sources understate what the models can do. Apple's local iPhone cleanup left a messy floor compared with the cloud tools.

2026-10-15

In "Nano Banana 2.1 vs GPT Image 2.5: Same Prompt, Same Product, 13 Rounds" (YouTube, a88Qf1kdIfo), the creator runs identical prompts through both image models and concludes that neither wins across the board. The right choice depends on the job. Nano Banana 2.1 takes photographic realism, layout design and overall aesthetic. It wins the editorial beauty portrait, which looks more realistic zoomed out even though it still has an AI look in close-up, and it wins the iPhone 13 front-camera selfie. GPT Image 2.5 wins on prompt adherence, UGC and cinematic looks, and product integrity in ads. It takes the candid street photo for following the prompt more closely, the cinematic film still for looking better as photography, and the 2009 low-quality camcorder frame by a wide margin, because Nano Banana pasted the camera specs into the image itself. Both models show a plastic-like look in some photoreal outputs. Product fidelity is the deciding test for commercial work. GPT Image kept the Vibe Fuel bottle's text, shape and design consistent, including in the old-way/new-way ad. Nano Banana rounded the rectangular bottle and distorted it. The creator argues that brands and clients will not accept a product that looks distorted or different from the real one, so GPT Image takes ads and creatives. Ads are not a clean sweep, though. One hard product case was a tie, and Nano Banana won the rating ad for its minimalist, consistent design despite slight distortion at the top of the bottle. The creator's tally also credits UGC to GPT Image overall, even though the selfie went to Nano Banana. The creator calls the judgments subjective, and this is the first of a planned series of side-by-side model comparisons.

2026-10-15

In "Higgsfield AI Marketing Studio - The AI Ad Generator I'm Actually Impressed By" (https://www.youtube.com/watch?v=E3AUM3XuH2I), the presenter demos Higgsfield Marketing Studio, which runs on SeaDance 2. It turns a product image or e-commerce link plus a consistent avatar into a finished ad in one of nine formats: UGC, tutorial, unboxing, hypermotion, product review, TV spot, wildcard, UGC virtual try-on and pro virtual try-on. The presenter says this work used to cost "15 to $40,000" and take weeks, and now takes one afternoon. They name solo founders, small brands, UGC creators and marketers who need 10 variations by Friday as the audience. An ad is defined by two slots, the product and the face selling it, plus a format, and the prompt is optional. A blank prompt produced the UGC perfume ad. A one-line descriptive prompt carried a specific detail, a dial turn that changes the weight, into the unboxing. The hypermotion ad used only the format name. Product intake is automatic: an uploaded image is analysed for name and description, or a product URL is scraped. The 40 preset avatars, and custom ones built from text or a photo in about 20 seconds, keep the same face across every format, which the presenter says is what makes the tool usable for real campaigns (compare Consistency as the trust signal in ad campaigns). Four demos serve as the proof: a UGC perfume ad, a dumbbell unboxing, a hypermotion soda shot and a pro virtual try-on. They are judged on product-identity fidelity (label, cap and dial accuracy, lip sync, natural hands, garment cut and fabric). The presenter calls the try-on the standout because likeness, garment and location were all correct in one shot, and says most try-on tools fail there. Duration is matched to format, with 8 seconds for UGC and 12 for the unboxing reveal. The demos use 9:16 at 720p, within advanced settings of seven aspect ratios and 4 to 15 seconds. Hypermotion drops the avatar and speech for pure product motion, generated in about 30 seconds from a two-word prompt. The perfume and soda products were themselves AI-generated images, and the dumbbell and clothing came from Amazon links. Finished clips can be dubbed into another language with lip sync or given a different voiceover. Work is organised into projects and a gallery with a recreate button. A separate "URL to ad" tool is named as the fully automatic path but is not shown. The presenter argues that cost per generation decides whether you test 10 ideas or two, and says Higgsfield is currently the cheapest place to run SeaDance 2. The video ends with a description link offering free credits with no payment. The cost, speed and "cheapest" claims are the presenter's own and are not independently verified. The video is a promotional hands-on demo, not a comparison.

2026-10-15

In "I tested viral AI 3D products" (Blender Guru, https://www.youtube.com/watch?v=qsmDdOi07fI), a 22-year Blender user and 3D-asset company founder replicates three viral AI-3D demos and grades each true, plausible or false. ChatGPT driving Blender through Python scripts is graded plausible. The mechanism is "legitimately insane", but his scene came out far simpler than OpenAI's, even after a six-hour, 20x-cost multi-agent attempt. He suspects the demo was micromanaged room by room, and he concludes AI assistance suits bounded tasks such as checking normals or writing a short geometry-nodes setup, and is weaker at judgment-heavy work like furnishing a house. The Tripo and Meshy mesh-generation ads are graded false, though he insists the technology itself is real. Tripo's 8K-texture claim could not be reproduced: the model had about 60% UV coverage against the 70-85% of a good game asset, so it delivered roughly 1K texel density at 8K performance cost. Meshy's 3D-print ad used a professional marketplace model (Cyberbee), and his own generated version failed in the slicer. Both companies conceded, and Meshy pulled four adverts. Mesh generation already substitutes for photogrammetry in film and TV, where blur doesn't matter at a distance. 3D-guided AI video (Blender blocks driving a video model, per a Higgsfield tutorial) is graded true. Reproducing it took many generations, colored faces on the guide block to fix character direction, an exact prompt specifying the number of ninjas, and hand-rewritten choreography. It handled no more than two ninjas at a time, and consistency over a long film would be frustrating. The recurring pattern is hidden human effort behind viral demos, and the published prompt alone didn't reproduce the result. On job loss, he cites the precedent of Cycles automating bounce-light placement without mass displacement, because saved time gets refilled (Parkinson's law) and falling barriers create new industries. He notes that AI differs because there is no end to the skills it can automate and it moves far faster. He predicts the future rewards "chefs" (vision, story, directing agents or humans, iterating faster) over "cooks" (paid to do one task in a preferred tool), without endorsing that outcome. He ends ambivalent: he fears artists quitting, and he also sees AI possibly letting him finally make a short film alone.

2026-10-15

A walkthrough at https://www.youtube.com/watch?v=3rDs6FhFoUQ shows a full AI-generated product commercial built on a laptop inside Kixal AI. It uses a repeatable three-step workflow. First, build and lock tested visual assets with Soul Cinema and GPT Image 2.0. Second, have a preloaded Claude skill turn the script into one connected shot list. Third, generate and iterate scenes in the video model (SeaArt/Citas 2.0). The creator argues the real skill is iteration, not a perfect first prompt: the finished ad is the best few seconds out of roughly 100 generations, cut together. On assets, a single reference image is not enough, and the model needs multi-angle product and character sheets or it hallucinates mid-scene. The creator recommends gray-background sheets with one face only, to avoid facial drift. Candidates should be tested in a simple moving scene before being locked, because a strong still can collapse in motion. Every edit degrades an asset toward an 'AI slop' look, so state changes such as sweat or a new outfit should be pre-generated as separate reference sheets, and outfits locked via a reference image so clips from different runs stay consistent. On the shot list, one connected document with a shared style-prefix block (and per-scene overrides) lets a single edit propagate to every prompt. Choreography must be spelled out move by move by name, and the music track can be fed in as an input so movement syncs to the beat. The creator calls the highest-leverage trick a generated spatial schematic/map attached to the prompt, which pins character and prop positions that text cannot and replaces brute-force '20 generations to fish out a few seconds'. The same pipeline reproduces a snorricam-style body-rig shot that would traditionally need a full-day shoot. Overall it frames AI ad production as an engineering-and-iteration problem, and the shot-list and schematic techniques are the most transferable parts. The specific tool stack (Kixal AI, Soul Cinema, Citas 2.0) is likely to date quickly.

2026-10-15

In 'How to Train a FLUX.2 LoRA with AI Toolkit' (a release-day walkthrough), the speaker argues that Flux 2 can be trained on day zero with AI Toolkit. Flux 2 is a 32B guidance-distilled model that does both text-to-image and image editing. It comes with a 24B vision-language text encoder (Mistral Small 3.1) and a new 32-channel autoencoder. Because one model does both jobs, a painter's style can be taught on plain captioned text-to-image data under a trigger word, then invoked on edit inputs with a prompt like 'make this picture [trigger] style'. No paired edit datasets are needed. Captions start with the trigger and describe only content, because anything left uncaptioned (style features, and clothing or hats if left out) binds to the trigger. Differential Output Preservation (DOP) pairs trigger captions with a preservation-class version such as 'photo'. The LoRA-off output of the base model becomes the target, so the style is learned as the difference from what the model already knows. In the speaker's results the style arrived quickly, complex scenes stayed coherent, and edits of a cat and a Mona Lisa kept fine paint-stroke detail. The speaker credits DOP for the retained comprehension and the 32-channel autoencoder for the detail. DOP alters every input, so it cannot be combined with cached text embeddings. The text encoder therefore stays resident, which raises VRAM use. Blank Prompt Preservation is the alternative that works with caching. The main cost is size. Both the transformer and the text encoder need float8 quantization, even on a 96 GB card, and training takes about 64 GB and does not fit on 48 GB. The run took 3,000 steps at roughly 6.4 s/iteration. Rank 32 gives a modest LoRA because the official implementation merges QKV into one layer, so most layers effectively need about 3x the rank. Training uses guidance 1, which is a no-op equal to CFG 0, and inference uses guidance of about 3. The speaker presents everything as an experiment. The dataset was heavily skewed toward white women, so a male subject became a woman and the speaker's dad lost his mustache. The suggested fixes are lowering the LoRA weight, rebalancing the data, or describing the subject in the prompt. The VRAM floor is unknown, ComfyUI support was unconfirmed at recording time, and the speaker is unsure instruction-style training is the right approach. Follow-ups are planned on character swapping and lower-VRAM training, including CPU offloading and accuracy recovery adapters.